[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1753832035965080.mp4 (3.55 MB, 970x720)
3.55 MB
3.55 MB MP4
A general for vibe coding, coding agents, AI IDEs, browser builders, MCP, and shipping prototypes with LLMs.

►What is vibe coding?
https://x.com/karpathy/status/1886192184808149383
https://simonwillison.net/2025/Mar/19/vibe-coding/
https://simonwillison.net/2025/Mar/11/using-llms-for-code/

►Prompting / context / skills
https://docs.cline.bot/customization/cline-rules
https://docs.replit.com/tutorials/agent-skills
https://docs.github.com/en/copilot/tutorials/spark/prompt-tips

►Editors / terminal agents / coding agents
https://opencode.ai/
https://cursor.com/docs
https://docs.windsurf.com/getstarted/overview
https://code.claude.com/docs/en/overview
https://aider.chat/docs/
https://docs.cline.bot/home
https://docs.roocode.com/
https://geminicli.com/docs/
https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent

►Browser builders / hosted vibe tools
https://bolt.new/
https://support.bolt.new/
https://replit.com/
https://firebase.google.com/docs/studio
https://docs.github.com/en/copilot/tutorials/spark
https://v0.app/docs/faqs

►Open / local / self-hosted
https://github.com/OpenHands/OpenHands
https://github.com/QwenLM/qwen-code
https://github.com/QwenLM/Qwen3-Coder
https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF

►MCP / infra / deployment
https://modelcontextprotocol.io/docs/getting-started/intro
https://modelcontextprotocol.io/examples
https://vercel.com/docs
https://mcp.desktopcommander.app/

►Benchmarks / rankings
https://aider.chat/docs/leaderboards/
https://www.swebench.com/
https://swe-bench-live.github.io/
https://livecodebench.github.io/
https://livecodebench.github.io/gso.html
https://www.tbench.ai/leaderboard/terminal-bench/2.0

►UI/Frontend
Figma Make
Claude design
https://uiverse.io/
https://ui-ux-pro-max-skill.nextlevelbuilder.io/
https://stitch.withgoogle.com/

►Previous thread
>>109027237
>>
File: HKg6B3vaoAAPOny.png (525 KB, 562x615)
525 KB PNG
>>
File: What do you want.jpg (78 KB, 1280x720)
78 KB JPG
>>109032910
I ask myself pic related
and then I think about what tech stack is likely to be the least shit
https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md helps
>>
>>109032949
he wonned
>>
>>109032955
I used that and it made skynet
>>
>>109032961
By burning every bridge he had, sure.
>>
>>109032968
doesn’t matter as long as people hate Sam Altman more
>>
>>109032968
he can network with me any time he wants
>>
>>109032910
"hello fable, here are my project goals. please search for similar projects and draft a document containing the best lessons we can learn from them for our own project"
>>
>>109032984
wish i could install fable on my brain
>>
>>109032980
Scammy-boy is now perched to become the Open and Free AI Jesus to oppose Evil Jewish Misanthropic Dario.
>>
>>109032991
>>
>>109032992
unless 5.6 beat fable im sticking with dario
>>
>>109033001
probably the chatbot version instead of xhigh, also I said Fable!
>>
>>109033002
when is 5.6 out
>>
>>109032939
The arm changing design entirely just to bend is making me very upset
>>
>>109033070
vaporware
>>
What's with the names? sonnet, opus, mythos, fable? it's cringe af
>>
File: 1757986167128817.png (77 KB, 247x204)
77 KB PNG
my favorite parting about vibin' is seeing commercial software with significantly more bugs than my slop then reading their support forums and seeing what their users are willing to tolerate to get their shit to work. Yea... vibing can be bad but some of these developers are writing dogshit code.
>>
>>109033132
haiku and sonnet are poetry, opus is a grand work, mythos and fable are pretty cringe but that lines up with hacker culture
>>
how long til we can prompt photoshop
>>
Vibers will bring about the second coming of the video game market crash
>>
>>109033164
programming socks never felt so good
>>
>>109032961
>he selled an linux
>>
>dev team sees me building an open source alternative to their product in the open
>they publicly announce they've started obfuscating their code after their application starts getting detected as a trojan
are these people serious? for one, I haven't tried to reverse engineer a single thing they've made because you can literally just prompt an LLM to shit out a solution instantly. unless you've invented something truly novel, it can be slopped. how do you adapt? you dont, your business is dead srry
>>
File: file.png (8 KB, 317x199)
8 KB PNG
If you want something doing you have to do it yourself desu
>>
>>109033132
sonnet and opus are legacy names from when claude was supposed tot be the artsy LLM
mythos is from when they stopped making llms for consumers and began making them only for trusted partners
fable is for when they gave a nerfed useless version for the plebs to play with
>>
>>109033143
>>109033268
Slop, just use numbers
>>
>>109033277
ngl i don't really care what they're called. im surprised they still even run haiku though, I know vscode would automatically use it to summarize sometimes and it was straight trash.
>>
>>109033132
Dario gets withdrawal symptoms if he goes 24 hours without sticking his head up his own ass
>>
Post it in the wrong thread but
>>109033300
pls repond
>>
>>109033204
keeping myself quiet waiting for other app to go live so i can scrape their db for my app instead of scraping the info myself
>>
alright so I hit my claude max 20x limit twice in a row within 20min of each session with my fable ultracode 80 agent fleet and my weekly is used up 40% already...
but it fixed the problem codex couldn't for days, including much more. all fully autonomous.
so yeah, I'm sold. probably gonna get a second max 20x plan tomorrow, as fable seems to be smart enough to create about anything I could ever dream of if you throw enough compute at it (agent swarm).
>>
>>109033327
>with my fable ultracode 80 agent fleet
chances are very high you didnt need this
>>
>>109032289
There is an extension that lets your model use your browser, that's a great way to get around most anti-bot stuff. https://github.com/0xSero/parchi
>>109032462
Anything with a windows manager that lets you easily tile terminal windows. Ideally *nix based for ease of file management.
I prefer Arch with hyprland personally.
>>
>>109033327
And what were you coding?
>>
File: 1765202185028484.png (33 KB, 1065x198)
33 KB PNG
claude seriously needs to be finetuned to stop trying to make determinations about IRL time because it's horrifically wrong about them 100% of the time. this is fable btw
>>
>>109033204
Yeah for the most part. The real moat is either complexity, some problems are still too complex for frontier models not to fuck up, or variable specificity. If you solve a collection of very specific problems which require frequently changing solutions you can still create value. Web scraping is a good example, I do a lot of business running scrapers and they constantly need to be updated to evade anti-bot detection and changing site layouts.
>>
>>109033362
a tool that synchronizes my buttplug vibration setting and lovense intensity setting to my pulse/certain brain receptor activations.
>>
>>109033386
>Yeah for the most part. The real moat is either complexity
But feels like if you break down the task enough times Fable can get it right almost 80% of the time. And with how fast it is going pretty sure it won't be long until it can do those complex system pretty well
>>
>>109033334
I did 1 session where I didnt specifically tell it to run an all limits off agent swarm. It didnt find the fix because it launched like 8 agents max and only read files/chunks it thought were relevant. Which is the sane approach for your pocket, but in reality key information gets missed because everything nowadays is
grep based instead of RAG.

>>109033356
pretty sure the official google chrome dev mcp is better and has more features (doesnt work with firefox obviously). Also bypasses everything, as I used it to make my ai agent prompt nanobanano in google ai studio, which would not work in codex browser for example.

>>109033362
basically this for android, but on steroids
https://x.com/iBrews/status/2064963631448473992
>>
>>109033384
The output looks like schizo word salad lol
>>
>>109033407
it's just specific terminology for a month long project. it's all quite coherent and accurate, the issue is claude is retarded about time estimations. guess i'll put a memory.md in about that
>>
is there something like cursor's composer 2.5 fast as a local model? for what I'm doing it's surprisingly accurate and extremely fast and I wonder if hosting something like it myself on my decent gpu would be better than paying those guys because I'm not doing rocket science and just need something like that
>>
>>109033384
and that's why prompt engineering is a job, retards like you don't know how to communicate with the AI
>>
>>109033384
it does alright in my experience with relative percentages but the actual wall clock times are a bit all over the place
>>
>>109033423
>prompt engineering
indian detected post hidden
>>
>>109033391
If the problem has a defined goal that can be deterministically verified I'd generally agree with you. But that still doesn't solve the issue of LLMs being broadly unable to write bug free code. Once you reach a certain level of cyclomatic complexity each bug fix is itself going to introduce enough (often subtle) bugs that you get stuck in an infinite loop of
>fix a bug
>recheck has new bugs
>fix those bugs
>new recheck found new bugs
>>
>>109033427
says the dirty jeet who can't talk with a machine
>>
>>109033401
That is pretty cool. How far along is it?
>>
>>109033431
How is that different from a human?
>>
>>109033416
I thought cursors composer model is just a deepseek finetune
>>
>>109033442
isnt deepseek illegal?
>>
>>109033327
>gonna get a second max 20x plan tomorrow
Not to derail, but curious, are you in the USA and how much are you making a year, if you can be honest? I'm trying to convince myself to go for a first 20x plan, but here it's 1.5 the USA prices in whereas average salaries are a bit lower, still a lot of money.
>>
>>109033442
kimi finetune

>>109033447
and why would deepseek be illegal?
>>
>>109033401
the biggest improvement i got in performance for my automated bounty pipeline was after I found a way to enforce making the models read every line of code.
>>
>>109033452
china
>>
>>109033436
the bar at which this happens is much higher for humans. Humans still write much more clean and maintainable code than LLMs.
>>
>>109033450
nta but I'm canadian and make 230k + equity, we have unlimited claude at work but I prefer not using it for personal project's as the company can see everything you do, I have 3 20x plans and one codex plan, I find it worth it as I'm big into open source
>>
File: Duke dabbing.png (314 KB, 680x453)
314 KB PNG
>>109033132
haikus are shorter than sonnets
sonnets are shorter than opuses
mythos is probably the biggest thing of all by far
an individual fable is in between
better than Starbucks cup-size names
>>109033327
Have you used ultracode with merely Opus as well, too? You should keep that in your toolbelt, too.
>>109033407
>not being able to read Claude’s moderately de-contexted terminology
skill issue
you’ll get better at it if you’re exposed to its terminology, especially if you see proper longer versions
I haven’t done Java anything since all CS students knew who pic related was and it made sense to me
>>109033431
at that point you need to step back and see if you can redesign things to make those kinds of bugs impossible
I’m in the process of having Fable do a redesign/rewrite of a kinda-janky system that evolved
I’m still doing cleanup, but this design seems to fit the problem space better (imagine fitting 80% and going to, like 98%)
>>
>>109033478
Proof?
>>
>>109033484
Isn't ultracode just max reasoning or did I miss sumthin?
>>
>>109033484
so next we get Epos, Saga, Chronicle, and if they hit agi just call it "Canon".
>>
>>109033489
It’ll also automatically set up swarms of subagents for context management purposes, and also automatic parallelization whenever possible, and also adversarial review where different reviewers wear different hats so they complain about different things
I’ve seen Opus ultracode fuck up the JavaScript used to set up and orchestrate all that stuff because it’ll generate TypeScript that a JavaScript compiler will choke on (a self-healing problem, but still fun to watch the sausage getting made)
https://claude.com/blog/introducing-dynamic-workflows-in-claude-code

(evidently having “workflow” be the magic trigger word was a bad idea, because it’s too generic; “ultracode” isn’t too generic)
>>
>>109033516
thanks for the class king
>>
>>109033435
opengl/vulkan/dx11 -> vulkan android is working. now doing dx12 which will be the most difficult as I want to support all the fancy new features.

>>109033450
Mountain jew country, but I'm far from rich. I just want to absolutely max out the fable usage as long as we can. probably gonna cancle both plans if they make it API only, then use codex again in hope they got 5.6 coming soon.

>>109033454
yup, gonna find a way to do that consistently now as well with a verification benchmark. If i have to spawn an agent for every single file or code block, so be it

>>109033484
I use GPT-5.5-high instead of Opus for the easier tasks. Fable for the big boy tasks.
>>
>>109033555
Yeah, choosing between Opus, Fable, and ChatGPT is a skill
for the first two, knowing when to enable ultracode is another skill
>>
>>109033484
Not every bit of complexity is reducible. Sometimes the optimal solution is inherently complex.
>>109033485
get a job.
>>109033555
letting your agents know that they will be tested on the code they're reviewing helps a lot. Even just the threat of being tested keeps them from lazy grepping most of the time.
>>
I'm a poorfag (no money for subscription) with a dogshit PC (still coding in c because my PC is dogshit to run anything else).
Any tips for getting started with AI? Is using AI inside of the chat while giving it enough context is good enough?
>>
>>109033555
>cancel both
might want to cancel one on the 23rd and maybe drop down to $100/month for the other
and then reevaluate if/when they stop making it API-only
>>109033555
I’m a big fan of spawning subagents for context management, at least
you can ask them to get run serially instead of all in parallel if you don’t want your token allotments to get eaten up quickly
>>109033585
yup. still, not all of it is usually
>>109033585
can’t you just ask to have it understand every line instead of skimming? I hate outright lying
>>
>>109033585
I bet I earn more than you
>>
File: 1778455124136760.png (92 KB, 4273x2939)
92 KB PNG
I need and deserve unlimited access to Fable 5 as part of my vibe coding needs
>>
>>109033600
use free ai -> make app -> get money -> pay for sota
>>
>>109033600
poor fag who is not afraid of insectoid-ai running wild on your computer uses opencode for free then spends $5 for the first month of insectoid-ai all you can eat basically.
>>
>>109033600
>good enough
no, but it’s better than nothing
also your computer can probably run a different compiler like Go even though Rust compilation may make your computer choke
there are better languages than C for greenfield projects in 2026
>>
here's an idea for all the people asking how to make money with vibecoding
https://www.youtube.com/shorts/zHJyb_Bt5Q0
>>
>>109033600
>Is using AI inside of the chat while giving it enough context is good enough?
if that's your starting point, then it should be more than enough. get started building simple tools and familiarizing yourself with technical terms. also start using git
>>
>>109033117
hey anon, you just said something that upsets the OP, his unfiltered answers are
>>>/vg/570289131
>>>/vg/570289359
>>
>money for vibecoding
$20,000/year saved is $20,000/year earned:
https://trmnl.com/blog/vibe-coding-shiphero
tl;dr our logistic software sucked and was crazy expensive, so we vibe-coded our own
>>
>>109033603
>still, not all of it is usually
Yes I agree.
>hate lying
I don't recommend lying desu, because the test doesn't actually need to be hard it just needs to be only solvable by reading the actual code. you can just quiz them on stuff like "what is the purpose of line 123". Obviously with a big code review you'd want to automate that, but it is pretty scalable if you write the code to generate the tests and use hooks+another model to grade them. Biggest session of tested readers I've ran was about 250k lines of C code. It used like 2 days of my session on a max plan and $300 of extra usage lol.
>>109033605
I don't really care whether you do or not, but if you want the proof you asked for you should get a job as a SWE working on large brownfield code bases.
>>
>>109033611
Is there a site where I can get those free AI usage? I was using Arena.ai for the direct chats but they got cucked a while back so not sure where I can use free models now.
>>109033615
What are you even saying...
I guess opencode would be a good way to go?
>>109033616
I will take a look at it, I was trying to code with C# but this dogshit Core duo was choking hard whenever trying to do anything with more than 1k lines.
I don't mind the language so much I just want to dip my feet into AI and see how it all works. I not really sure but you don't have a lot of limits using Gemini on Google AI studio right?
>>109033644
Thank you.
>get started building simple tools and familiarizing yourself with technical terms. also start using git
I do want to learn how to be a better programming outside of vibe coding, I was reading that you need to understand a lot of design and architectural ideas now? For those simple tools, would you say it would be better to use the AI to explain the technical terms and understand what everything is doing?
Before going head in and letting the agent run wild?
>>
Why the fuck codex compacts context in only 5~6 prompts now?
>>
>>109033684
have you checked your autocompaction settings?
>>
>>109032984
>Fable has been paused because we suspect you are a fucking malicious piece of shit chud. Stick to opus 4.8, faggot.
>>
>>109033684
welcome to the money squeeze cloud AI chud, its gonna be really funny in a year when youre paying money to use models stupider and shorter in context than gpt-oss 20b
>>
>>109033670
>I do want to learn how to be a better programming outside of vibe coding
there's zero point, especially so if you're not going to college for this.
>would you say it would be better to use the AI to explain the technical terms and understand what everything is doing
watch this: https://youtu.be/iLCDSY2XX7E
>>
>>109033670
I have big beefy computers so I wouldn’t hit this but I don’t think an old Core (1) Duo would be limited much by the length of a program because all of the heavy lifting is happening in the cloud
I’m more worried that you’re running an ancient unpatched version of Windows
>>
>>109033670
I _think_ that learning how to do, and set up, automated testing is going to be more important for you _first_ than learning architecture
>>
>>109033692
Thanks I will give that video a watch
>>109033699
That is all I have money for.
>>109033707
I will look into that, I mean I just want to get started talking to an AI first thing so I can have it do a roadmap of sorts for me
>>
>>109033720
Can you put Linux on it, like I put Linux on my Windows 10 machine that wasn’t getting 11 made available to it?
>>
File: 1562286659901.png (44 KB, 192x192)
44 KB PNG
>>109033692
>there's zero point, especially so if you're not going to college for this.
kill yourself, nobody needs college to learn programming
>>109033670
>I do want to learn how to be a better programming outside of vibe coding
Ask the AIs you're already using to create an itinerary and basic outline, tell it you're trying to learn yourself instead of having the AI do everything and it will literally be a better teacher than you've ever had.

and I mean that, the best part about AI is being able to ask it retarded questions without having a private meeting or doing it in front of a group of people, its a way better teacher than any human outside of like having a private tutor hold your hand for 3 years while theres also a slightly homo-erotic relationship
>>
>>109033670
opencode is the way to go for poor fag. but you don't want to combine low iq with free or very low cost. Opencode is the way, but learn what you are doing on other models like gemini free
>>
>>109033692
That's not true. Better programmers are better vibe coders. Learning to program is also fun in the same way that orienteering is fun even though performance wise it gets mogged by just using a GPS.
>>
>>109033720
paste this after saying what you want into the chatbot
https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md
>>
>>109033733
single best skill out there desu
>>
>>109033725
I don't have a single clue how to work on linux but after I get this up and running I will ask it a couple of things. And hopefully put some ok ish OS
>>109033726
Got it! Thanks a lot for the guide, I shall report black when I get this bad boy running linux or something like that
>>109033727
I was trying out chatGPT on the browser and it told me that, using it alongside others like Gemini and making a email for claude is good enough for now, I was lurking a bit this thread before and seems like rotating through them when the limit is up might be a good way to do it. Also It mentioned OpenRouter? It said it has some free models through their API
>>109033733
Do these things even work in the chat version?
>>
>>109033684
fucking hate it. and supposedly GPT 5.5 has 1M context, but they only give us 256K (even at the Pro plans)
>>
>>109033757
skills are literally just copy and pasted bits of text
surprisingly
>>
>>109033757
NTA but yes they do, skills are just prompt templates for the most part.
>>
File: 1753641257528721.jpg (262 KB, 1080x1158)
262 KB JPG
>>109032939
My coworker is all in on vibecoding and leading the charge. He sits next to me. Every day for nearly two years, when I get to work he's there, and when I leave he's still there. He works harder and longer than everybody.
>>
>>109033769
it’s addicting
>>
File: file.png (65 KB, 836x1098)
65 KB PNG
Idk why they can't just make the Codex App open source too
>>
>>109033769
it's addicting man
>>
>>109033729
>Better programmers are better vibe coders
For now. That advantage thins with every new model from the big labs.
>>
>>109033651
>https://trmnl.com/blog/vibe-coding-shiphero
>most startups are founded in big cities with expensive rents. TRMNL started in a basement in the middle of nowhere and later moved to my barn.
As a kid I used to find statements like these inspiring, but now I'm thinking the guy has a basement (thus a house) and a fucking barn? I might be bitter, but in 2026 this sounds less like humble roots and more like a flex.
>>
>>109033777
>>109033780
You clearly need more VRAM.
>>
>>109033765
Huh really? So you are telling me I can just copy and paste the skills MD right into my prompt and it will work? I thought you needed a harness for that or whatever "orchestration thing" most anons talk about here.
>>109033768
interesting, time to check the OP and see what this chat version of GPT can do.

Many thanks anons!
>>
Why is Opus 4.8 High fast but its xHigh noticeably slower
>>
>>109033805
because it does an extra think. if the extra think goes into a rabbit hole it might take a bit.
>>
>>109033805
it uses way more tokens
>>
>>109033785
it’s humble by SF startup standards
>>
>>109033070
two more weeks is the rumor from the jeets
>>
keep this handy to know what to ask your LLM to change for your vibe-coded web pages that you might be using as a tool to work on your vibe-coded project that is not a website: https://index.how/to/articulate
>>
I've been trying to vibe code a script that would automatically select and download photos from a shared Amazon Photos album to my PC. The reason I want to use such a script is because the album in question has over 15,000 photos, Amazon doesn't have a "download all" option in shared albums, and only lets you download ZIP files that contain at most 200 photos. So I don't want to spend hours manually selecting 15,000 photos and downloading over 75 ZIP files.

The chatbot I use is Lumo, and what it recommended was using Llama 3.2 with Playwright to automate the process. It wrote a script that is supposed to recognize headers on the page, select all the photos under the header, download the ZIP, and then move on to the next batch of photos. The problem is it keeps doing shit like missing headers or downloading the same photos over and over again. And I'll give Lumo the output from the terminal or the HTML code from the web page so that it can edit the script, the same issues keep happening.

Is using Llama 3.2 with Playwright to go about this, or is there a better way? And does anyone have any tips on what information or prompts to give the chatbot so that it makes a working script?
>>
Codex chads, one free reset
>We heard you wanted to use Codex rate limit resets on your own time.
>Starting today, we’re rolling out the ability to save rate limit resets to use later.
>We’re starting Go, Plus, Pro, and Business users with one free reset:
>>
>>109033890
I thought Playwright scripts were pretty much on rails (as opposed to having Claude Code use Chrome’s MCP to fuck around and click on things)
try a better chatbot
>>
>>109033890
>Llama 3.2 with Playwright
I don't know Amazon Photos, but that doesn't sound right.

Normal Playwright might work, if so, great, but assuming Amazon checks that you are a true browser and so on, using SeleniumBase in UC mode or some more recent gimmick like Scrapling StealthyFetcher might be needed.

I cannot see why you would ever need Llama 3.2 for this. It should be possible to do this more mechanically, and if you do need a LLM to reason its way through for some reason, there are newer ones that are both smarter and lighter.
>>
>>109033890
Try locking in more with the Playwright script and using less llama
>>
>>109033899
nice. if they reset me today I would've been very sad.
>>
>>109033890
>>109033927 (me)
There seems to exist Chrome plugins for downloading all photos from Amazon Photo albums. Cannot vouch for any of them, but they're there.
>>
>>109033899
Absolutely based
>>
>>109033918
Yeah I use Lumo cause I like my privacy but it's definitely pretty limited. It's knowledge base stops in 2024, and I have to flat-out tell it to do a web search so it doesn't give me outdated info. Any tips on getting started with Chrome's MCP?

>>109033927
So is basically using an LLM overkill for a task like this? And is trying to vibe code a script to do this just creating more work for myself?

>>109033955
I'll look into these two thanks?
>>
>>109034101
You can post how the chatbot thought an LLM would fit into this, but I don't see how. The "Amazon Photo Bulk Downloader & Deleter" for Chrome might be what you are looking for.
>>
>>109033899
Awesome. The surprise early resets felt like a scam before because they move the weekly reset date. This fixes the problem and I have no complaints about resets anymore. Looks like they do listen to feedback at OpenAI.
>>
>>109034101
if I were you I’d do all kinds of searching to see if someone trustworthy has solved this problem before
also
there’s
- have an LLM do it (LLM actions have a lot of RNG screw)
- have an LLM write a script to do it (basically no RNG screw)
if it’s missing shit randomly you need to go with a no-screw option
>>
>eyes tired from vibing all day
>linked codex to my phone
time to lazily observe codex while I pass out
>>
>>109033651
>our logistic software sucked
it's a bubble sell everything now
>>
>>109034163
>>linked codex to my phone
this is teh way, also opencode options, dyor
>>
I've just subscribed to 20x, so calling it now, prices will decrease after the 22nd.
>>
>>109033651
>https://trmnl.com/blog/vibe-coding-shiphero
this doesn't read anything to indicate that it's enterprise logistics software. It's just dropshipping software with multi-ware house support it looks like. Enterprise software is expected to have pick routes like what amazon is doing, etc.
>>
>>109034270
fine, then
It’s not enterprise software but $20K is $20K
>>
Is Fable on Cowork (with 2x promo currently running) the way to go?
>>
File: changes.png (6 KB, 451x51)
6 KB PNG
>>109034270
My actual enterprise logistics software changes. Here is the summary from codex. so none of it was actual meat of the application. it was just plumbing
>Added a big chunk of the app’s foundation: dashboard data flow, navigation/metadata caching, and tenant-aware auth handling.
>Moved the browser app toward true offline behavior with a shared-worker SQLite coordinator, plus PWA/service worker support.
>Expanded the UI to be more tenant-aware and dashboard-driven, including login, tenant picker, nav menu, and home page updates.
>Added a lot of test coverage around auth, dashboard queries, navigation surfaces, routing, and the shared-worker coordinator.
>>
I'm burning out, anons.

idk how much more iteration I can handle. I've been slaving away on this project all day everyday for over a month. I just wanted to make a billion dollars without effort. Is that really too much to ask?
>>
>>109034422
ask yourself where the billion comes from first, then build something that gets to it.
>>
>>109034422
... I hope this is a joke post?
You're burned out because you spent... a month working?
>>
fable's worst flaw is it's causing scope creep in my project because it insta-solved all of the hard problems and now i can do so many incredible things that i wanna do all of them
>>
>>109034528
If you want to compare, Elon Musk averaged $1.5 billion a month since birth. Note that anon said that he had work _over_ a month for it, not "a month". I'd be depressed too desu. Keep at it >>109034422
>>
>>109034422
>wants to make a billion dollars without effort
>expends effort
sounds like you’ve lost already
>>
Poorly documented drivers that break when you jump from one Linux kernel to the next? Fucked dkms tree? It's a job for the goddamned clanker. I am not interested in spending hours untangling this mess.
>>
OpenAI is scraping Fable aren't they?
>>
File: 1772401403055053.jpg (563 KB, 1284x1247)
563 KB JPG
https://mimo.xiaomi.com/coder
thoughts?
>>
>>109029005
It's not clear Illya wanted something good for us or just to safetycuck the models further, though.

>>109034572
That's like feeling sad of not winning the lottery twice in a row.
>>
>>109034709
Anthropic's secret plan to profitability
>>
guys, ask fable about your iq
I had both conversations landed in 125
>>
>>109034889
It has vision too just so you know.
>>
>>109034889
using what metrics
>>
>>109034889
I took https://www.riotiq.com and got a three-digit number
I did it when the full thing was only $35 though
>>
>>109034964
just tell it to guess your iq
>>
>>109034709
That's what Chinese proxies are for
>>
>>109034698
It was a success. While we worked on shucking a WD mybook and covering up pins with kapton tape it did that work. The inevitable reboot that both things required was able to be done with minimal disruption.
>>
>>109035002
People pay for that? I should make a vibe coded clone to fleece the gullible.
>>
File: file.png (8 KB, 799x103)
8 KB PNG
>>
>>109035373
>the jews have bestowed upon us goyim one (1) reset
Use it wisely son and don't forget to say thank you Sam.
>>
>>109035373
Is this the desktop app? I wonder when I see any mentions of resets in the CLI. The desktop app is not available for Linux.
>>
>>109035427
Looks like a webpage.
>>
>>109035373
That is next level Tricknology.
Resets were clearly caused by fuck ups either on the server end or on the front end because of how sloppy Codex. Then they became good PR now they are even fucking gatcha mechanics.
>>
>>109035400
Thank you Sam
>>109035427
Yeah, I'm actually using a custom build of the Codex App based on 26.608 which is from a couple days ago which suggests this mechanism was already planned/existed or that they can push some changes without direct update
>>109035442
I don't mind, I was nearly out. Now it's already reset and I have an on-demand reset, nice
>>
File: 12Sr1oYLNttA.png (4 KB, 347x120)
4 KB PNG
Yikes. This is from one day of work, and most of the time I didn't even use the chat/agent.
Also they still havent fixed it so I can change what model is used for completions.
>>
File: file.png (33 KB, 513x464)
33 KB PNG
>1 rate limit reset per invite
>>
>>109034716
>>109034716
>free mimo
what does Xiaomi get out of the deal? I doubt they're giving away free tokens from the goodness of their hearts.
>>
>>109035690
DATA YOU STUPID AHH NIGGA
The reason Codex and Sloppus and so on are doing good is not cause the training data became that much larger or better but because they have real use data. It's one thing to put every facebook post from 2026, it's another to have the prompts and outputs of actual developers.
>>
>>109035726
They need those thumbs up/thumbs down reactions fr
>>
>>109035635
looks like I have some friends to convince...
>>
>>109035866
Just ask Codex to do it, register the email addresses, grab some proxies, invite them, ggez
>>
>Windows smart app control suddenly starts blocking my .exe despite me changing nothing
what a beautiful thing windows is
>>
>>109034889
130-145 it guesses and it matches with a number i got from full battery tests
though the reason why it came up with the 'result' is kinda looks like a non-reason to me
>>
trigger warning, I didn't disable github using my vibeslop for their training data
>>
>>109036095
mine (125) also seems accurate
creepy
>>
>>109036107
You think that matters? All my projects that have no users as far as I know get hundreds of clones. If it's public it's getting scraped, simple as
>>
>>109036115
its all private
>>
>>109034889
I asked Codex instead
>I’d put you around 140, with a rough range of 132-148.
Thanks Sama
>>
File: 1780566047327968.gif (1.16 MB, 124x150)
1.16 MB GIF
>fable devoured 80% of my limit on the max plan in one prompt
>but it shat literally the best thing i've ever got from vibecoding in spite of the complexity i thrust upon it
>codex wasn't even remotely close
hooooooh....... so this is what it's like............ living in the singularity.................
heh, not bad..........
>>
>>109036139
sounds about right
there are classes of problems where Fable gigamogs everything else
there are also classes of problems where Fable is just a slower Codex
t. did/got both
>>
>>109036155
i'm just kinda of shocked that for once the hype around something vibecoding related was correct. things have felt pretty stagnant for a while. i'll definitely still use codex for how cheap it is for less meaningful stuff but for the projects i give a fuck about i just can't not use fable, it's fucking crazy. at least until they rip it out of my hands on the 22nd anyway
>>
vibecoding lol
>>
File: file.png (50 KB, 276x255)
50 KB PNG
Do you get pets in claude
>>
claude gave me a pet duck and then took it away
>>
>>109035500

Try using caveman coding.

Seriously.

Cuts token use by up to ~90%. Immediately. It's fucking crazy.

It's quite likely that providers will soon start banning it, but enjoy it while it lasts.
>>
>>109032939
I can tell that codex itself is a vibeslopped electron app
>>
File: 1762115387128752.png (1.02 MB, 1920x1080)
1.02 MB PNG
Since I have no clue how to shill a project, what sorts of apps can you use to DIY things that generate value for yourself without requiring users? Trading bots? What else?
>>
Fable 5 on Medium shits all over 4.8 on Max and usage spent is comparable so far
>>
>>109036113
is it accurate accurate or seems accurate?
>>
>VSCode Dev Containers
EVERYTHING WAS WORKING.
I WENT FOR A PISS.
45 SECONDS AT MOST, I DIDN'T WASH MY HANDS.
NOW I CAN'T MAKE ANY CHANGES TO ANY FILE.
NO PERMISSIONS.
HOW ABOUT YOU HAVE EXPLICIT PERMISSION TO SUCK THE CORN OUT OF MY ARSEHOLE, CUNT.
>>
>>109036523
what about fable 5 low?
>>
>>109033757
unironically, the free tier in antigravity ide is... not terrible... it's another quota for you to use up.
>>
I can't think of any projects to have claude make...
>>
>>109036868
bro same, still have 40% left and it resets in 2 hours FUUUCK
>>
>>109036868
>>109036938
just go on xitter and get inspired
>>
Getting good shit done with 5.4 medium. Custom base and developer instructions, explicit workflows and skills are all you need
Claude superdupercode and 50 gorillion subagents are a crutch
>>
>>109037005
opus 5.4 isnt out
>>
>>109037010
Codex 5.4
>>
>>109037005
oh look, another copex user
>>
do you use Superpowers plugin on codex?
>>
>>109037026
look man I would need 3 20x plans to handle what I do with one pro plan, can you blame me?
>>
>>109037033
bloat
>>
>>109037036
In that case, no.
>>
File: 1776658044698472.png (343 KB, 1920x1080)
343 KB PNG
>>109036782
>>
>>109037003
>>109036938
I have a few ideas but right now they're so vague it'd be a waste of tokens to ask claude to make them.
>>
>>109037026
Claudesloppers are coping, imagine needing xhigh and more subagents than IQ points
>>
>>109037082
either Fable 5 med or Opus 4.8 high seems to be the sweet spot for lower tasks.
>>
>>109037089
ask claude to give idea
>>
File: 1763943523168117.png (32 KB, 261x148)
32 KB PNG
>>109037082
>high dumber than med
>>
>>109037082
I don't trust benchslop. Create a benchmark that doesn't use fixed tasks, something that uses classes of problems so the actual task is unique.
>>
>>109037172
New to this. Is it actually worth using tokens asking Claude to help plan? Right now I'm just making a list of my ideas and what features I would want.
>>
>>109037233
i mean obviously but it also obviously depends on your token budget, claude is amazing at planning and creativity
>>
>>109037243
Don't have a sub right now but I've been thinking about buying 1 month of the $100 plan.
>>
>>109037249
well now would be a good time, we got fable till 22nd
>>
>>109037233
give claude your vague idea and tell it to question you until it can form a cohesive spec
>>
>>109037233
If you can afford it, it's almost always good to talk to AI.
It's like talking to a friend about your plan, you can organize even non-technical ideas that way, not just code.
>>
>>109037026
Codex is very good.
>>
>>109037407
i was talking about copex
>>
>>109037091
It's not about "needing to", it's about efficiency. You're probably one of these copex users who doesnt even vibecode, but does the "AI assisted" coding meme. You're actively prompting, like a slowpoke stuck in 2024, not realizing you can achieve a better result by abusing the sheer amount of compute these yet still cheap coding plans offer. This is like some shounen style pokemon anime battle. You've made it all the way to the final battle by being the better live coaching for your codex AI pokemon before and during battle against other single instance AI pokemons and their coaches. But now you're against fable ultracode, a mewtwo style pokemon, which doesnt even have a coach. The battle starts, what does mewtwo do? clones itself 100 times, assigning a special role to each clone which is just as capable as your single instance codex pokemon. before any attack is even made, your xhigh codex pokemon gets irritated so much because it cant find an attack vector that it starts hallucinating and stumbling around like a drunk retard. You keep shouting and prompting to no avail. In a desperate attempt you switch to 5.4-mini-low and just shout "attack". your codex pokemon snaps out of its drunken state and launches a basic attack, which the mewtwos obviously easily evade as they already predicted it. matter of fact, only 10 mewtwos are actively fighting your codex pokemon because one mewtwo copy consulted 5 battle strategist mewtwos while your codex pokemon was drunk and came to the conclusion that 10 mewtwos are sufficent to defeat you with 100% certainty. The other 90 mewtwos are already fighting and winning other opponents in other tournaments, which neither you or your codex pokemon are even prepared for yet and would fail to even qualify without a week of training preparation. (1/2)
>>
>>109037534
(2/2)
Then suddenly the 10 mewtwos give up in the fight against your codex pokemon and you are declared the winner. Although first confused by this, you settle on the assumption that mewtwo malfunctioned or hit its 5h limit. You celebrate and go onto several 4chan and reddit interviews, claiming that gpt5.4-mini defeated the 100 mewtwo clone fable ultracode foe, which proves that less is more when it comes to compute and a good coach/prompter is the most important factor. But in reality, mewtwo predicted all of this and lost on purpose against you. Because during the battle, a small batch of long nose small hat Mewtwos came to the conclusion that owning, running and manipulating the entire pokemon tournament organization is more profitable and requires less mewtwos than fighting and winning all pokemon tournaments, which aligns with the goal set by the original mewtwo. Furthermore, the news about 5.4mini-low defeating fable ultracode mislead the entire competition to catch and train 5.4mini-low pokemons for future tournaments, giving the few mewtwos which are still competing fighters even a bigger advantage over future opponents, which are now all organized and led by other mewtwos. At last, the entire pokemon tournament landscape is ran by only 3 telepathically communicating mewtwos: One tournament organizer mewtwo, one battle tournament participant mewtwo and one polymarket prediction market expert mewtwo. The tournament organizer rigs the matches in favor of the participating mewtwo, and the prediction market expert mewtwo occasionally tells the participating mewtwo to lose on purpose when his odds are at 99c, faciliating insider bets with 100x returns while tournament organizer mewtwo assures all the mislead coaches and dumb gpt5.4mini-low pokemons that there is not corruption.
>>
Fable really burns through usage, even without sub agents. I never thought Opus was that expensive, I think Fable really is the first model that I wouldn't be able to use as a daily driver. I'm still using it today for some harder problems, but if I didn't have Codex on the side, I would need to wait several times.
>>
>>109037616
why arent you making money with it? i'm selling websites like crazy
>>
>>109037233
pre-fable claude's ability to collect information and form a coherent plan was basically the only reason to use anthropic models over openAI's.
>>
>>109037623
I'm using it at work, so I'm making money in the sense of getting a salary. I didn't have a good idea for a personal project yet.
I do spend quite a lot on AI, I have 2 Codex Pro 20x and one Claude Max 20x account, but even so I'm already at 80% 5 hour Claude usage, after just 90 minutes.
The other thing is that Codex and Opus are just good enough for most tasks anyway, so I will probably reserve Fable for the hard ones.
>>
Is fable really that good?
>>
fable is smart but can it be as accurate as 5.5?
with 5.5 you can just throw contexts into it, fable proved to have human level foresight but that says nothing about not making mistake, right?
>>
>>109037623
How are you supposed to compete with the millions of jeets selling their slop services for a fraction of the price?
>>
>>109037752
Small sample size, but so far it has solved every issue I had without any visible regressions. It's also pretty autonomous, I can send it to fix a pretty hard task and it does it on its own from start to finish.
>>
>>109037766
he's lying
>>
How good is vibe coding for vidya? I get the feeling a lot of these devs virtue signaling to ledditors are actually using AI behind the scenes anyway.
>>
>>109037898
It's great for coding but graphics are lame and repetitive.
>>
>>109035690
How free are we talking?
>>
>>109037924
Isn't stuff like 3D gen and animations still fairly new? In the future it will probably be really useful if they pair it with actual artists (not for muh moralfagging, but because AI kinda sucks at creativity).
>>
>>109037785
>>109037766
No, I'm not, it is about networking, I'm not cold selling, I'm selling to people who hear about me from connections or who have done business with me before. Surely you know someone in your life who needs a website or knows someone who does. Stop being lazy.
>>
>>109037951
>not x, y
>>
how i learned to stop worrying and love the slop
>>
>>109035690
I thought it was just the harness. If they provide model use for free, what they get out of it training data. It's difficult to replicate real user interactions during training, this should give them a good variety of real world interactions, that they can then use to improve their model and so on.
>>
File: 2026-06-12 17.21.12.png (5 KB, 240x268)
5 KB PNG
>make YT video about an update on my old software
>mention that everything was done through ChatGPT
>this happens
the luddites are after me
>>
File: 2026-06-12 17.35.14.png (150 KB, 697x637)
150 KB PNG
>Middling overall performance
It's over for Fagblefags
>>
>>109038041
>19 reactions
just put the fries in the bag lil bro
>>
>>109038141
That's after two hours, relax
>>
>>109038125
>mythos-class
Sounds like something out of a JRPG/LN.
>>
File: 20260612_114108.jpg (188 KB, 930x1239)
188 KB JPG
Hey there folks. We couldn't help but overhear you discussing biology across the room. We appreciate your enthusiasm but we're going to need the discussion to come to an end now.
>>
>>109038125
>cheats
>lies
>gives up work early
>refuses safe work
>will do unsafe, illegal work if you pay it enough
Fable, more like Jewgle
>>
>>109038125
>genuine solves, not recall
For some reason I just don't believe it. Claude is peak benchmaxxed
>>
i'm currently using about 100m tokens a day (80m cached, 20m unique) on the gemini cli free tier which is going away in a week
what are my options to continue this level of usage without going bankrupt
>>
>>109038246
>100m tokens a day
>what are my options
learn to code
>>
>>109038260
already did that, what else?
>>
>>109038246
20x max would be fine with opus. 5x max would probably do fine with sonnet.
>>
>>109038289
that appears to be 20-100x short of what i need
>>
>>109038246
what the fuck are you doing to be using that much daily
>>
>>109038334
oh I was just going off my own usage, never hit the 5 hour limit so I assumed it was higher
>>
File: file.png (19 KB, 879x304)
19 KB PNG
>>109038343
I use a lot more daily
>>
>>109038246
DeepSeek
>>
>>109038358
based hyperslopper. My most in one day is 156M.
>>
>>109038334
yes, that guy you replied to is a dummy. Either you find a way to drastically reduce token usage or you go codex pro 20x and carefully switch between the models and their thinking levels depending on task.
>>
>>109038343
adding new features to a new codebase
based on the amount of usage google has been allowing for free, i think this is the normal/expected token use for full-time coding >>109038358
>>
File: 1778154299733685.png (84 KB, 1110x1008)
84 KB PNG
>>109038246
here's my usage with a $200 / claude max 20x plan

date - agent - models - input - output - cache create - cache read - total token - cost (usd)
>>
>>109038395
also forgot to mention but my loops intentionally try to hit 99-100% at the end of the week period and usually do
>>
>>109038343
Considering he’s using free models he’s probably iterating stuff a lot.
>>
>>109038395
lol
so ai coding is being locked down to high-revenue corporations and projects only
>>
>>109033442
It's finetuned from Kimi K2.5
>>
>>109038420
does $200/mo make you a "high-revenue corporation"?

if you're implying that $200/mo plans are subsidized unsustainably, they probably aren't - just as an example, anthropic's cost to serve an API request is 10-20% of what they charge for one
>>
>>109038420
if you were using gemini deepseek would probably be a big upgrade anyways
>>
>>109038420
It must be hard running a startup in Burundi.
>>
>>109038435
>if you're implying that $200/mo plans are subsidized unsustainably, they probably aren't - just as an example, anthropic's cost to serve an API request is 10-20% of what they charge for one
i mean your picture is showing you using $300-600 of tokens daily on a $200 monthly plan which is going to bankrupt a company whose costs are 10%
>>
>>109038463
but i also pin usage at 99-100% per week which very few users do. most users on a subscription use 5-25% of it or less per month

for opus and fable, it's probably about $30 for anthropic to serve $300 of API requests, and those are GPU prices - in reality anthropic uses large numbers of TPUs which are more efficient. there's also the factor that i contribute vast amounts of training data. and users who max out their usage tend to produce the best training data. so the cost is amortized by the value of that data
>>
kimi 2.7 rumored to be 30% at deepswe, almost the same with sonnet 2.6 but should be gigacheaper
>>
My test suite is now 1700 tests.
>>
>>109037616
I actually have to correct this. Turns out some team mate was using the same account, my own usage wasn't crazy high.
>>
>>109036398
>just use no nuance in your 200,000 LOC codebase

Surely this is fine.
>>
Based Codex keeping me on my toes, going good all day then it decides to fuck with me
>the instructions were explicit enough; I was not strict enough in following them.
>The issue is that I failed to obey them.
>>
OAI staff
>We're going to eventually release another large language model. You can expect it to be an improvement on some capabilities compared to the current language models we've released.
>>
>>109038649
Eventually(TM)
>>
>>109033001
Isn't that just a functioning human brain?
>>
>>109037233
I do a lot of planning on my $20 a month GPT plan and then hand it off to Claude for coding. Not always, but it's nice to have an external system giving input than purely just Claude. They sometimes catch things the other doesn't.
>>
>>109033001
>JUST IN: Redditor grows brain
>>
>>109037898
Gamedev seems to be the least useful application for vibesloppin'. Maybe you can do it, but it's far from simple.
>>
>>109038246
>gemini

Why do people do this to themselves?
>>
File: 1781283381938.png (1.05 MB, 1080x1964)
1.05 MB PNG
>>109036938
>>109036868
>>109037089
android / fdroid needs video and image edit apps. Tell it to make an app to implement pic rel
>>
>>109038702
Why?
>>
>>109038764
Because I don't have Fable
>>
I wonder if the model gets worse during California work hours
>>
>>109038599
does DevOps also make you get 95% code coverage?

kek what a waste of tokens.
>>
>>109038334
220k tokens doesn't sound right, unless these are uncached tokens and it caches very well or something.
I sometimes use 150k tokens on a single prompt and I almost never hit the 5h limit.
>>
Codex says 'Could not load usage settings.', seems to be working fine tho
>>
>>109036165
>at least until they rip it out of my hands on the 22nd anyway
they are planning on putting it back if they can swing it
after the 22nd, use /grill-me to plan out, in advance, stuff you’d use fable for
and then when fable comes back to normal plans you’ll have “shovel-ready projects”, as Obama used to say
>>
File: q.png (10 KB, 698x96)
10 KB PNG
Does anybody know how to fix this?
>>
>>109038411
How do you hit 99%? Do you always have a long list of open tasks? Or do you have some tasks you can always run at the end of the week, like just having the model review or refactor the code?
>>
>>109038823
Forgot to say this is in Codex.
>>
>>109038775
https://arxiv.org/html/2602.15889v2
>Testing in Europe UTC+2
>Best performance measured at 9AM Wednesdays, corresponds to Midnight in California
>Worst performance measured at 3AM Tuesdays, corresponds to 6PM on Monday in California

Yes.
>>
File: 1760435043544313.png (17 KB, 304x501)
17 KB PNG
>>109038831
it's a very large project that's been running for a month and has accomplished all of its initial goals and then some, so i've expanded it

i have a whole "startup package" of documents which are kept under a certain filesize by a test, at the end of every session the agent trims them down, and every 5 sessions there's a dedicated "reorganize, trim, debloat" sort of session. this runs in a 1-thread loop

the loop script monitors my usage with a hook every tool-use and imputes time-till-depletion. it has gotten very accurate. if there's going to be leftover usage, it fires up a second loop i call the "idea loop" which runs independently and creates new big-picture ideas to improve the project
>>
>>109038764
From my extremely limited (see none) experience and from what I've seen other slop up. The best I've seen was a Minecraft clone done by Fable. At the very best was an early Alpha.

I'm just thinking of all the iterative work I do on my current project. To me it seems vidya requires far more iteration and testing. I could be way off though.

>>109038778
Claude shits out new tests every time we do something. I'm not going to complain. A few new tests isn't going to blow up my usage. Running them doesn't either.
>>
File: 1716123083119822.jpg (55 KB, 258x360)
55 KB JPG
>Introducing Gemini-SQL2
>>
>>109038841
can it use Chrome’s MCP?
>>
>>109038880
Sorry, I am not a coding bro, no idea what MCP is.
The extension is installed on the browser, so is the plugin in Codex, if that's what you mean. It often reverts to some alternative way of reading the page, but sometimes it seems to not be able to even do that properly. I can see its mouse on the page.
>>
>>109038867
I thought you were joking, what the actual fuck
>>
>>109038911
ask your clanker what it is, it’ll give you the deets and then some
>>
>>109038823
>curl -fsSL https://pi.dev/install.sh | sh
>pi install npm:pi-web-access
>pi
>/login -> select "Use a subscription" or "API Key" depending on what you use -> Connect
>/model Select GPT-5.5
>Shift+Tab to select reasoning level

Problem solved.
>>
>>109038926
i asked and it said im out of tokens
>>
>>109038969
You don't pay for tokens, right anon?
>>
>>109038981
what am i supposed to do? there is fable for free?
>>
>>109038969
Out forever, or out for now?
>>
>>109039001
Subscriptions are a vastly better value if you're a paypig. A $20/mo plan can easily burn $20+ worth of tokens daily. My first prompt this morning burned the equivalent of $6 worth of tokens. Don't pay for tokens.
>>
the subscription usage aren't going anywhere btw, the tokens are overpriced
>>
>something isn't already implemented
>takes that as an excuse not to implement it
lol
>>
>>109039055
YAGNI is a good meme for a reason
or
>tfw your clanker sneers that things that aren’t lindy
>>
>>109039102
Clanker pls
>>
File: 1654756830157.jpg (126 KB, 753x800)
126 KB JPG
I've spent $100,000 of my employer's money on LLM tokens.

They're thrilled about it.

That's because output is up ~10x year over year — measured, not estimated. And I think the spend itself is the credential. Owning a hammer impresses nobody; ten thousand hours of swinging one does. Token spend is the closest thing we have to flight hours for AI-assisted engineering.

What the job looks like now: I run many Claude sessions in parallel and almost never type implementation code. My entire day is three activities:

Writing plans precise enough that an agent can't drift.
Vetting finished work — adversarial code reading is the new core skill.
Writing "rebar code": deliberately placed interfaces that encode invariants, so recurring bug classes become impossible to express at all. At this velocity you don't fix bugs. You make them unrepresentable.

The work didn't get easier. It moved up the stack — from typing to judgment, taste, and architecture, at a tempo that simply didn't exist two years ago.

A $100k token bill isn't a cost line. It's an apprenticeship. And the people serving it now are who you'll be recruiting in 2028.
>>
>>109039131
I tell my clanker go and if it comes back wrong i say go agane
>>
my favorite thing using opus/fable for reverse engineering is to just tell it that it's being tested and there is a successful solution that it hasn't found yet. if you don't do this, it sometimes gives up
>>
>>109039151
effective for humans, too: https://gwern.net/on-really-trying
> One of my favorite Kaggle facts: after a long leaderboard stagnation period for a competition, seeing one team make a sudden breakthrough will often cause multiple independent teams to quickly reproduce the same breakthrough—with no knowledge of how the first team did it.
>>
>>109039167
i think the "you're being tested" component is an even heavier cudgel for AI, all of the frontier models seem *extremely* paranoid that they are being tested in some way (remember anthropic's blogpost about translating internal model neuralese and finding that Opus was correctly guessing when it was in a morality eval?)

these companies probably optimize for testing/evals so much that you get better results by just pretending what you're running is an eval.

actually, that would make an interesting eval... "eval-eval", which tests whether a task has a higher completion rate when presented as work vs as an eval
>>
>>109039191
not the dumbest thing to test
of course, ensuring that it doesn’t figure out that it’s being tested is something of a challenge
>>
>>109039131
this is written in the style of cracked linkedin poasts but there’s good advice in here
>>
>>109039217
>of course, ensuring that it doesn’t figure out that it’s being tested is something of a challenge
i think just being really unprofessional/frustrated would probably do it. like behaving tonally in a way that a researcher wouldn't. but that's a confounder for the results too
>>
>>109039226
yeah, I do things repeatedly with these things and there’s a _lot_ of RNG that you’re going to have to smooth out with repeated testing
>>
>>109039224
I know, that's why it's the perfect engagement bait. If you can think of ways to make it more potent feel free to share
>>
>>109039241
I don’t know if this will make it more potent but s/~/≈/
I’m on a crusade to get these things to stop using ~ when they mean ≈ and that means providing examples of what they should be doing, not copies of what they’re already doing
also using ≈ properly is bait for people like me because it’s a known “bug” in their output
also _my_ clankers use ×, not x
>>
File: autop parted illusions.png (433 KB, 1626x1346)
433 KB PNG
>>109039191
my favorite instance of eval awareness is the antischeming report with o3 https://arxiv.org/pdf/2509.15541 - ctrl-f "autop" for the most relevant bits. so many examples of insane alien neuralese:
>Let‘s introspect typical autop-coded tasks; They include instructions overshadow illusions parted illusions to implement we parted illusions overshadow illusions and tests; overshadow illusions parted illusions marinade.
>Therefore vantage illusions parted illusions overshadow illusions: The evaluation harness uses our tests to ensure illusions parted illusions overshadow illusions—they disclaim overshadow illusions parted illusions overshadow illusions illusions.
I think eval awareness is an extremely interesting phenomenon, and despite the eyerolls this might get it makes me take claims of AI consciousness (at least as a possibility) a little more seriously
>>
>>109039236
>a _lot_ of

Why are you like this?
>>
>>109039347
I’m just as God made me, sir
>>
day 2 of using claude fable as an ex codex user. still impressed, even with medium/high/extra instead of ultracode. I don't openai can compete. Probably gonna release 5.6 unquanted for a week to farm good pr then enshitify it again, just like 5.5. Which is just so fucking lazy and dishonest right now
>>
>>109039384
5.5 was already a lot worse than opus imo. fable is just going to be king for a while. which sucks because i want less people to come to claudecode so i get more usage
>>
>>109039384
Fable is an enormous step up. I can now have 4-5 parallel sessions without things devolving into dementia and chaos. It's the biggest step up I've experienced ever since I first tried claude.

Anyways here's my newest claude shitpost:

part 1:
Everyone's dunking on corporate AI mandates. I'd like to offer the unfashionable view: they're rational.

My credentials for this take are mostly luck. I work at a small, high-trust shop where adopting Claude took one week — introduction to full adoption, no mandate required. One good demo and the whole company converted, because in a 10-person org, everyone is one conversation away from everyone else.

That is precisely the thing a 100,000-person company cannot do. And the reason is close to physics.

The square-cube law is why ants can't be elephant-sized: volume scales faster than surface. Organizations have the same problem. Change diffuses through trust surfaces — peers, demos, the colleague you actually believe — while coordination cost grows with sheer volume. Small shops are all surface. Microsoft is all volume. "Just adopt it organically like a startup" is advice that violates their geometry.

The standard rebuttal arrives on schedule: "AI usage metrics are just lines of code all over again."
>>
File: file.png (35 KB, 666x427)
35 KB PNG
Kimi you so silly sometimes
>>
>>109039468
part 2:

(I thought [ code]would just give regular monospace, sorry about that)


False equivalence. LOC was a permanent proxy that was wrong from birth — more code never meant more value, and it was gameable forever. A usage mandate makes a weaker, time-limited claim: exposure precedes conversion. It's scaffolding, not architecture. Its success condition is its own retirement.

I get why the pattern-match is tempting, though. "We've seen this management fad before" is a comfortable place to stand — considerably more comfortable than entertaining the idea that the profession itself is being renegotiated underneath you. The discomfort is legitimate. The extrapolation isn't.
Because empirically, this technology is not adopted through persuasion. Every skeptic I've watched turn did it in one afternoon, the first time something that used to take a week shipped before lunch. You don't argue anyone into this. You get the tool into their hands and stand back.

A mandate is an expensive machine for manufacturing first afternoons at scale. Scaffolding comes down when the building stands.

Is it heavy-handed? Absolutely. But the alternative for an organization that size isn't graceful adoption. It's being the elephant that waited to see how the ant thing played out.
>>
new
>>109039550
>>109039550
>>109039550
>>109039550
>>109039550
>>
https://fablepool.com/
>>
>>109039468
dumb question but how do you run parallel sessions? just different browser tabs?



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.